Participants
Patients encountered in any of the inpatient, outpatient or emergency room settings were considered eligible if they were adults between 18 and 100 years of age with documented AF or AFL and taking one of the following DOACs: apixaban, rivaroxaban, or dabigatran. Any classification of AF was accepted (i.e. paroxysmal, persistent, permanent), regardless of the control strategy (i.e. rhythm, rate) and irrespective of previous procedural interventions (i.e. ablation, cardioversion). Patients with valvular AF (i.e. in the setting of rheumatic mitral stenosis or prosthetic valves) were excluded. Other exclusion criteria were a history of venous thromboembolic disease (VTE) such as deep vein thrombosis (DVT) or pulmonary embolism (PE), in order to exclude patients with competing reasons for anticoagulation. Patients who were taking different antiplatelets such as P2Y12 inhibitors (i.e. clopidogrel, prasugrel or ticagrelor) were also excluded. The International Classification of Diseases, Ninth and Tenth Revisions, Clinical Modification (ICD-9-CM & ICD-10-CM) codes were used to identify the study patients. The cohort was divided into two groups: individuals taking ASA in addition to a DOAC (exposed group) and patients taking a DOAC without ASA (unexposed group). The indication for ASA use and the dosage was not assessed. All subjects were observed for a minimum of 2 years. The outcomes of interest were identified by querying hospital readmission diagnoses, inpatient diagnoses, discharge diagnoses, and active problem list using ICD-9-CM and ICD-10-CM codes.
Variables
The primary outcome is the composite major adverse cardiac events (MACE) defined as (1) ischemic cerebrovascular events including stroke and transient ischemic attack, (2) systemic embolism to any vascular territory outside the central nervous system, (3) and acute coronary syndromes (ACS) including unstable angina, non-ST elevation and ST elevation myocardial infarctions. Secondary outcomes are all cause mortality and bleeding, defined as any bleeding event leading to hospital presentation or admission; the severity of bleeding was not addressed as all events were considered severe if they prompted hospital presentation. Only the first event was analyzed and patients who experienced subsequent events were censored after experiencing any of the above outcomes.
The variables assessed included: patient age, gender, and race, in addition to multiple comorbidities that can have an effect on the risk of developing cardiovascular disease and bleeding. We also calculated a CHADS-VASc score for each patient (a validated clinical prediction tool for estimating the risk of stroke in non-rheumatic atrial fibrillation) and the HASBLED score for each patient (a validated scoring system developed to assess 1-year risk of major bleeding in patients taking anticoagulation with atrial fibrillation).
Bias
All patients identified in Beaumont’s healthcare database who met the inclusion criteria were included in the study in an attempt to minimize selection bias. Propensity scores were calculated for baseline characteristics and used to inversely weigh all observations in an attempt to achieve balance in the treatment groups and minimize confounders.
In an effort to minimize information (measurement) bias, automated reports of patient data and outcomes were generated by an individual who was not involved in the study protocol or statistical analysis. Covariates, outcomes, and baseline characteristics were obtained in a standardized fashion without knowledge of the patient groups. Moreover, regular meetings with the data collectors were held to ensure variables were obtained in a consistent fashion, thus minimizing inter-observer variability. Additionally, our biostatistician was not involved in the study design and data collection or interpretation. Researcher bias was limited via strict adherence to the study protocol. Finally, the impact of residual confounding was minimized by adjusted analysis for known confounders; however, the potential for unidentified or unknown confounders exists.
Statistical methods
Differences in baseline characteristics between the two treatment groups (DOAC+ASA and DOAC only) were compared using the χ2 test for categorical variables and the Student unpaired t test for continuous variables, as appropriate.
Before analyzing outcomes, a propensity score was calculated for each patient in the analysis dataset. Propensity score was defined as the estimated probability of being “treated” (which for this study means having index treatment of “DOAC+ASA”) as a function of covariates. The following covariates were included in the calculation: sex, race, age, tobacco use, body mass index (BMI), CHADS-VASc score, history of anemia, coronary artery disease (CAD), cancer, congestive heart failure (CHF), chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), diabetes mellitus (DM), gastrointestinal (GI) bleed, myocardial infarction (MI), obstructive sleep apnea (OSA), peptic ulcer disease (PUD), stroke, peripheral vascular disease, baseline use of non-steroidal anti-inflammatory drugs (NSAID), protein pump inhibitors (PPI), statins, angiotensin converting enzyme inhibitors (ACEi), and beta blockers.
Propensity score was then used to balance the treatment groups in terms of covariate distributions by weighting each observation by the inverse probability of treatment. Additionally, because there were a few observations with extremely large weights, we standardized the weights by the actual (sample) proportion of treated. Weighting results in a synthetic sample in which the distribution of baseline covariates is independent of treatment. Once balance in covariates was achieved, weighted data was used for subsequent analyses. A Cox proportional hazards model was employed to estimate hazard ratios for each of the three outcomes (MACE, bleeding, and death). Treatment was included in all models, and adjusted for sex, race, age, tobacco use, body mass index (BMI), CHADS-VASc score, history of anemia, coronary artery disease (CAD), cancer, congestive heart failure (CHF), chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), diabetes mellitus (DM), gastrointestinal (GI) bleed, myocardial infarction (MI), obstructive sleep apnea (OSA), peptic ulcer disease (PUD), stroke, peripheral vascular disease, baseline use of non-steroidal anti-inflammatory drugs (NSAID), protein pump inhibitors (PPI), statins, angiotensin converting enzyme inhibitors (ACEi), and beta blockers. HASBLED scores were not included in the calculation of propensity scores or in the adjusted models because aspirin use automatically adds a point to the score, thus none of the subjects in the exposed group would have had a score of zero.
Time-to-event curves comparing treatment groups for MACE, bleeding and death were created using predicted probabilities of event-free survival from the adjusted Cox regression models. Number Needed to Harm (NNH) was calculated using the predicted MACE rates from the weighted and adjusted Cox proportional hazards models.
All analyses were conducted using SAS version 9.4, (SAS Institute, Cary, NC). Statistical significance was assumed at a p-value < 0.05. All tests were 2-sided.